receiving, by data processing hardware, an indicator matrix composed of at least four indicators for a product of at least one store, the indicator matrix based on historical operational transaction data of the at least one store, wherein the historical operational transaction data includes at least one of inventory deduction data, inventory increase data, inventory scrap data, abnormal consumption data, profit data, or loss data, and wherein the at least four indicators include weighted mean absolute percentage error WMAPE, days of inventory DOI, demand fulfillment rate DFR and markout rate MOR of inventory;

automatically adjusting, by the data processing hardware, the indicator matrix;

determining, by the data processing hardware, for each indicator of the at least four indicators, an expected indicator range and a baseline of the product in a plurality of stores;

providing, by the data processing hardware, external data contemporaneous with the historical operational transaction data inte to a target feature extraction model, wherein the target feature extraction model processes the external data to extract a plurality of target features that cause the indicator matrix to satisfy the expected indicator range for at least one indicator of the at least four indicators;

providing, by the data processing hardware, the plurality of target features and the historical operational transaction data to a machine learning model trained to predict a demand for the at least one store and provide replenishment suggestions for the at least one store based on the plurality of target features and the historical operational transaction data; and

causing, by the data processing hardware, automatic replenishment of the at least one store based on the machine learning model.